Dynamic Prediction of Day Mortality in Patients With Trauma Using a Hybrid Neural Network Model: Model Development...

The trajectory of a patient with trauma is often complex and nonlinear. Real-time estimation of the mortality risk from prehospital care to discharge is critical for point-of-care decision-making and for benchmarking the quality of care. Conventional risk assessment systems in trauma are simple and data-sparse, leaving potential for harvesting available data for personalized risk assessments accounting for developing patient states.

This study aims to create an AI risk prediction model for 30-day mortality in patients with trauma capable of performing predictions at any time point through treatment phases from prehospital to discharge.

Data on pre- and in-hospital care of patients with trauma treated in Denmark, Capital Region between 2017 and 2024 were used. Demographic and comorbidity-specific variables were structured as tabular data. Temporal data including vitals, laboratory test results, and medications were structured as sequences with temporally determined dynamic bin sizes. Trajectories were censored before the outcome event. Across continuous input types, scaling and normalization parameters were fitted on observed values only, and missing values were handled through zero-imputation with auxiliary indicator channels enabling the model to distinguish observed from absent measurements. Missing values across categorical input types were mapped to a reserved token with a learned embedding. The model architecture combines both tabular and sequential input data. A unified input is processed by multiple layers of transformer encoders before being passed into a neural network for binary classification. Model performance was assessed using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) on a holdout dataset comprising 20% of the total data.

A total of 9496 patients were included. The model achieved an AUROC of 0.962 (95% CI 0.930-0.994) and an AUPRC of 0.655 (95% CI 0.545-0.765) predicting 30-day mortality in the holdout evaluation set using full-length trajectories, comprising 1829 patients. In active-cohort evaluation, the model achieved an AUROC of 0.905 (95% CI 0.856-0.953) at 1 hour from first patient contact, demonstrating early discriminative capability from the prehospital phase onward. The model outperformed both Revised Trauma Score (mean ΔAUROC +0.297; 50/54 time points significant) and Trauma and Injury Severity Score (mean ΔAUROC +0.169; 46/54 time points significant).

We designed a dynamic, automated model that allows risk prediction at any point in time during nonlinear trajectory of the patient with trauma. This study demonstrates that a hybrid neural network model, trained on automatically extracted electronic health record data, can predict 30-day all-cause mortality in patients with trauma with strong discrimination across the care trajectory. The model could be used as decision support for triage, patient deterioration alerts, bedside decision-making, and family counseling. These findings support the feasibility of sequential modeling for trauma risk prediction.